DeepSeek’s V4-Flash-0731 lands with agentic and coding boosts
DeepSeek has pushed its open-weight Mixture-of-Experts model into a new phase. The team quietly pushed DeepSeek-V4-Flash-0731 to Hugging Face and promoted its V4-Flash API to public beta on July 31, 2026. The headline is simple: the same 284-billion-parameter architecture and size as the preview build, but with fresh post-training that lifts agentic and coding performance without touching the underlying design.
A model tuned for agents and code
The upgrade arrives as a checkpoint that ships with the DSpark speculative-decoding module, mirroring the structure of the separate DSpark variant. DeepSeek reports gains across agentic and code-centric benchmarks, including Terminal Bench 2.1 and NL2Repo, where the new build outperforms both the preview V4-Flash and V4-Pro in several cases. Code Agent tasks were evaluated with a minimal mode of the DeepSeek Harness, and the company notes that scores can shift with different harness configurations. Internal test sets DSBench-FullStack and DSBench-Hard show the model’s added strength in full-stack scenarios.
How to use it today
Deployment comes in two flavors. On the API side, almost anyone can spin up the model: DeepSeek’s pricing page lists $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens, capped at 2,500 concurrent requests. That’s roughly a third of the V4-Pro output price and puts agent loops within reach for seed-stage startups and indie developers. Self-hosting remains more demanding: the weights are MIT-licensed and ungated, but every expert remains resident in memory even though only 13 billion parameters activate per token. DeepSeek’s vLLM example serves the model on a single 4×GB300 node, while Unsloth’s dynamic GGUF builds land at 162 GB in lossless 8-bit and 103 GB in 3-bit, requiring roughly 110 GB of combined RAM plus VRAM.
Inside the unchanged architecture
The technical report confirms V4-Flash is still a 284B MoE with 13B activated per token, 1M-token context, and hybrid attention combining CSA and HCA. Routing uses the first three MoE layers with hash routing, then switches to 256 routed experts per layer with six firing per token. Efficiency claims from the report—27% of single-token FLOPs and 10% of KV cache versus DeepSeek-V3.2 at 1M context—refer to V4-Pro, not the Flash variant.
Why it matters
This release lowers the bar for high-quality agentic and coding inference without forcing users to adopt new hardware or rewrite their stacks. For startups and indie teams, the API pricing makes it realistic to run multi-step agent loops at scale. Enterprises that need full control can self-host, but only with substantial memory and compute budgets. The bigger picture is a maturing open ecosystem where performance gains come from training tweaks rather than architectural overhauls, keeping the door open for broader experimentation.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.